Worse RANSAC results in version 0.15.2

[cattaneod]

Dear all,

I'm using RANSAC based on features (in my case, features extracted with a DNN) with the function registration_ransac_based_on_feature_matching, and I noticed that after updating open3d to the newest version 0.15.2, the results were worse than before.

So, to double check, I tested the same code on the same dataset with different versions of open3d. I'm aware that there's some randomness involved in RANSAC, however the results achieved with 0.15.2 are constantly worse (in the range of 3x-5x larger errors) compared to previous versions. Versions 0.12.0, 0.13.0, and 0.14.2 all achieve better results.